feat(format): skip the project load when paths are given

[tripleaceme]

Description

Closes #6025.

sqlmesh format already accepted positional paths and never loaded state, but with paths it still loaded every model in the project and then filtered them down with Path.samefile. That load is wasted: formatting pretty-prints a file's own text and needs nothing from the rest of the project graph.

Formatting a file needs three things — the file's text, the config of the project that owns its path, and the dialect and formatting flag. The first two come from config_for_path, which resolves a config from a path alone, and the last two are read off the file's own MODEL/AUDIT header. So with paths the project is no longer loaded at all. format now joins lint in scoping its own load: with no paths it loads the project itself, which also leaves the web and LSP callers untouched, since those contexts are already loaded and _loaded short-circuits the call.

Whether a file is a model is now decided by parsing its header rather than by membership in the loaded project. That keeps macros/*.sql and other non-model SQL a no-op and ignores .py model paths, per the acceptance criteria.

Both routes now collect FormatTargets and share one format-and-write loop, which also drops the samefile call that ran once per model in the project.

One behavioural note

A standalone audit is formatted when selected by path, while the project-wide pass still skips it. That is because Context.format iterates chain(self._models.values(), self._audits.values()) and an AUDIT (..., standalone true) is loaded into self._standalone_audits, a separate dict — so standalone audits aren't formatted today at all. I raised this on the issue: the acceptance criteria say to format "SQL models or standalone audits", which assumes behaviour that doesn't currently exist. The header-based check treats both kinds of AUDIT the same, so the path route picks them up.

I've left the project-wide route exactly as it was rather than quietly fixing the gap there. test_format_paths_formats_standalone_audits pins both halves so the difference is explicit and easy to change in either direction once you've decided.

Test Plan

Ten tests in tests/core/test_format.py:

• test_format_paths_does_not_load_project — spies on Context.load, asserts it's never called and that the selected model is still formatted while an unselected one is untouched
• test_format_paths_output_matches_loaded_format — formats the same project twice, once through each route, and asserts the files come out byte-identical
• test_format_paths_ignores_non_model_sql / test_format_paths_ignores_python_models
• test_format_paths_honors_formatting_false, ..._model_defaults_formatting_false, ..._string_model_defaults_formatting
• test_format_paths_check_reports_unformatted
• test_format_without_paths_still_loads
• test_format_paths_formats_standalone_audits

Each "file is untouched" assertion is paired with a file that must change, so none of them can pass merely because nothing was formatted.

Two in tests/cli/test_cli.py covering the CLI wiring in both directions.

pytest tests/core/test_format.py tests/cli/test_cli.py tests/lsp tests/web
215 passed, 3 skipped

The three test_dlt_* failures are pre-existing on main in my environment (ModuleNotFoundError: dlt).

Checklist

• I have run make style and fixed any issues
• I have added tests for my changes (if applicable)
• All existing tests pass (make fast-test)
• My commits are signed off (git commit -s) per the DCO

[tripleaceme]

@cmgoffena13 — closes #6025. With paths given, the project is no longer loaded at all: formatting a file only needs its own text, the config that owns its path, and the dialect and formatting flag off its own MODEL/AUDIT header.

One thing I'd like your call on, also noted on the issue. The acceptance criteria say to format a path if it is "a SQL model or standalone audit", but standalone audits aren't formatted today at all — Context.format iterates chain(self._models.values(), self._audits.values()), and an AUDIT (..., standalone true) loads into self._standalone_audits, a separate dict.

I've left the project-wide route exactly as it is and let the path route pick them up, with a test pinning both halves so it's a small change in either direction once you decide.

Zero checks, pending workflow approval.

[tripleaceme]

Heads up @cmgoffena13 — merging #6060 put this one into conflict too, so I've resolved it in c08310b rather than leave it for you to trip over. It's MERGEABLE again.

Only tests/cli/test_cli.py conflicted, and only because both branches appended tests to the end of the file, so both blocks are kept.

sqlmesh/cli/main.py merged cleanly and I checked both changes survive rather than trusting it: the OPTIONAL_LOCAL_COMMANDS gating for test --local from #6060, and the ("lint", "format") scoped load from this branch. They sit in different parts of the group callback and don't interact. Confirmed afterwards that format <path> still skips Context.load, format with no paths still loads it, and test --local still runs.

pytest tests/cli/test_cli.py tests/core/test_format.py tests/lsp tests/web
222 passed, 2 skipped

The test_dlt_* failures are pre-existing on main in my environment (ModuleNotFoundError: dlt).

Still open on this one whenever you get to it: the standalone-audit question from the description. Context.format iterates chain(self._models.values(), self._audits.values()), so an AUDIT (..., standalone true) is never formatted today. I've left the project-wide route as-is and let the path route pick them up, with a test pinning both halves so it's a small change either way.